PRAXIST
Autonomous research system that evolves solutions across agent generations
Emerged finds new products worth a look, daily — what’s new today
A rare piece of agent-assisted research you can audit end to end: frozen protocols, dated cohorts, open code, and the same board in front of you as the solver.
An independent project trying to play Drop7 well enough to average a million points a game, publishing every step. The best candidate so far, n-tuple lookup tables inside a four-deep search, averaged 516,155 points over 512 unseen games against 382,567 for a hand-written leaf — a screening result, the site says, not a qualification. The MIT repo holds TypeScript and C++ engines, a benchmark, frozen protocols and a results ledger. You can play in the browser and put your score on the same leaderboard as the policies, where the top human sits at 1,597,898.
More like this
SoL-Pi
NVIDIA's extension that cuts coding-agent token use roughly in half
Hops Run
A browser racer where the rival cars are models answering one call per move
jevman
An arcade maze turned into a latency test for decision models
SWE-sweep
A benchmark where agents get only a repo and must find the bugs themselves
no_human
Open-source agent that takes a ticket to a reviewed pull request, with proof